Once in a while, I do Kernel programming in kernel space. I was told that atomic operations uses LOCK instruction in x86 and hence these are not good to use. I was told that incrementing variable under spin locks is a better option than using atomic_inc.

Based on number of instructions that gets executed in atomic_inc function is very small compared to instructions it takes to do lock, increment variables and release lock. Hence I expect using atomic_inc is better option from performance perspective. My simple experiment also proved this. But my colleague says that when there is any PCI traffic such as Ethernet traffic, then one would see the atomic_inc is not relatively good.

Any comments?

Thanks
Ravi

Well, spin-locks use those same atomic instructions in most architectures.

Code for simplicity and for the general case. Remember that your solution is expected to work equally well on all architectures. It's highly unlikely that the microseconds you may actually save will be noticeable in practice. But the hours you spend trying to troubleshoot a weird timing-bug will be.

Thank you for the quick response.

I agree that many non-x86 kernels have similar implementation between spin locks and atomic operations. It is very clear that atomic operations are better than incrementing a variable under locks in non-x6 (Example: Power PC) environments. My question was specific to x86 based Linux where LOCK instruction is used in atomic operations.

Thanks
Ravi